Starstruk Logo MarkStarstruk

v0.14.0

Comfy-Org/ComfyUIv0.14.0Feb 17, 2026by github-actions[bot]

AI Summary

Major release introducing Node Replacement API, training with proper offloading, improved dynamic_vram handling, and new API nodes for Magnific Upscalers and Tencent 3D tools.

Key Highlights

  • Implemented Node Replacement API
  • Added training with proper offloading support
  • Fixed dynamic_vram issues with fp8 and lora workflows
  • Added Magnific Upscalers API nodes
  • Implemented NAG (Normalized Attention Guidance) on all Flux-based models

New Features

  • Node Replacement API
  • Training with offloading
  • Magnific Upscalers support
  • Tencent ModelTo3DUV, 3DTextureEdit, 3DParts nodes
  • Bria RMBG nodes
  • VideoSlice node
  • Left padding support for LTXAV text encoder
  • Qwen 2512 ControlNet (Fun ControlNet) support
  • 3D models viewing in outputs window

Full Release Notes

## What's Changed
* mp: dont deep-clone objects from model_options by @rattus128 in https://github.com/Comfy-Org/ComfyUI/pull/12382
* sd: delay VAE dtype archive until after override by @rattus128 in https://github.com/Comfy-Org/ComfyUI/pull/12388
* dynamic_vram: ops: Fix vanilla-fp8 loaded lora quality by @rattus128 in https://github.com/Comfy-Org/ComfyUI/pull/12390
* Fix 3D models not showing in outputs window by @guill in https://github.com/Comfy-Org/ComfyUI/pull/12381
* Add a VideoSlice node by @AustinMroz in https://github.com/Comfy-Org/ComfyUI/pull/12107
* [Trainer] training with proper offloading by @KohakuBlueleaf in https://github.com/Comfy-Org/ComfyUI/pull/12189
* Make built in lora training work on anima. by @comfyanonymous in https://github.com/Comfy-Org/ComfyUI/pull/12402
* fix(api-nodes): retry on connection errors during polling instead of aborting by @bigcat88 in https://github.com/Comfy-Org/ComfyUI/pull/12393
* Dispatch desktop auto-bump when a ComfyUI release is published by @benceruleanlu in https://github.com/Comfy-Org/ComfyUI/pull/12398
* [API Nodes] enable Magnific Upscalers by @bigcat88 in https://github.com/Comfy-Org/ComfyUI/pull/12179
* dynamic_vram: Fix windows Aimdo crash + Fix LLM performance by @rattus128 in https://github.com/Comfy-Org/ComfyUI/pull/12408
* ace15: Use dynamic_vram friendly trange by @rattus128 in https://github.com/Comfy-Org/ComfyUI/pull/12409
* model_patcher: guard against none model_dtype by @rattus128 in https://github.com/Comfy-Org/ComfyUI/pull/12410
* Update ace15.py to allow min_p sampling by @askmyteapot in https://github.com/Comfy-Org/ComfyUI/pull/12373
* Add a tip for common error. by @comfyanonymous in https://github.com/Comfy-Org/ComfyUI/pull/12414
* fix(api-nodes): add separate retry budget for 429 rate limit responses by @bigcat88 in https://github.com/Comfy-Org/ComfyUI/pull/12421
* ModelPatcherDynamic: force load non leaf weights by @rattus128 in https://github.com/Comfy-Org/ComfyUI/pull/12433
* llama: use a more efficient rope implementation by @rattus128 in https://github.com/Comfy-Org/ComfyUI/pull/12434
* Update command to install AMD stable linux pytorch. by @comfyanonymous in https://github.com/Comfy-Org/ComfyUI/pull/12437
* dynamic_vram: Training fixes by @rattus128 in https://github.com/Comfy-Org/ComfyUI/pull/12442
* Use torch RMSNorm for flux models and refactor hunyuan video code. by @comfyanonymous in https://github.com/Comfy-Org/ComfyUI/pull/12432
* Support generating attention masks for left padded text encoders. by @comfyanonymous in https://github.com/Comfy-Org/ComfyUI/pull/12454
* Fix some custom nodes. by @comfyanonymous in https://github.com/Comfy-Org/ComfyUI/pull/12455
* Add left padding to LTXAV text encoder. by @comfyanonymous in https://github.com/Comfy-Org/ComfyUI/pull/12456
* Add working Qwen 2512 ControlNet (Fun ControlNet) support by @krigeta in https://github.com/Comfy-Org/ComfyUI/pull/12359
* Patch frontend to 1.38.14 (from 1.38.13) by @christian-byrne in https://github.com/Comfy-Org/ComfyUI/pull/12469
* Remove unsafe pickle loading code that was used on pytorch older than 2.4 by @comfyanonymous in https://github.com/Comfy-Org/ComfyUI/pull/12473
* chore(api-nodes): remove "gpt-4o" model by @bigcat88 in https://github.com/Comfy-Org/ComfyUI/pull/12467
* Node Replacement API by @Kosinkadink in https://github.com/Comfy-Org/ComfyUI/pull/12014
* feat(api-nodes-Tencent): add ModelTo3DUV, 3DTextureEdit, 3DParts nodes by @bigcat88 in https://github.com/Comfy-Org/ComfyUI/pull/12428
* feat(api-nodes): add Bria RMBG nodes by @bigcat88 in https://github.com/Comfy-Org/ComfyUI/pull/12465
* Fix lora Extraction in offload conditions (+ dynamic_vram mode) by @rattus128 in https://github.com/Comfy-Org/ComfyUI/pull/12479
* MPDynamic: force load flux img_in weight (Fixes flux1 canny+depth lora crash) by @rattus128 in https://github.com/Comfy-Org/ComfyUI/pull/12446
* Remove workaround for old pytorch. by @comfyanonymous in https://github.com/Comfy-Org/ComfyUI/pull/12480
* Update workflow templates to v0.8.42 by @comfyui-wiki in https://github.com/Comfy-Org/ComfyUI/pull/12491
* Remove code to support RMSNorm on old pytorch. by @comfyanonymous in https://github.com/Comfy-Org/ComfyUI/pull/12499
* gitignore venv* by @alexheretic in https://github.com/Comfy-Org/ComfyUI/pull/12431
* Allow control_after_generate to be type ControlAfterGenerate in v3 schema by @Kosinkadink in https://github.com/Comfy-Org/ComfyUI/pull/12187
* Implement NAG on all the models based on the Flux code. by @comfyanonymous in https://github.com/Comfy-Org/ComfyUI/pull/12500
* Fix anima preprocess text embeds not using right inference dtype. by @comfyanonymous in https://github.com/Comfy-Org/ComfyUI/pull/12501

## New Contributors
* @askmyteapot made their first contribution in https://github.com/Comfy-Org/ComfyUI/pull/12373
* @krigeta made their first contribution in https://github.com/Comfy-Org/ComfyUI/pull/12359

**Full Changelog**: https://github.com/Comfy-Org/ComfyUI/compare/v0.13.0...v0.14.0